license: apache-2.0
base_model:
- YuYu1015/YuYu1015-Ornith-1.0-35B-abliterated
base_model_relation: quantized
pipeline_tag: text-generation
tags: - gguf
- llama.cpp
- qwen3.5
- moe
- gated-deltanet
- reasoning
- abliterated
- uncensored
- imatrix
language: - en
- zh
YuYu1015-Ornith-1.0-35B-abliterated-GGUF
GGUF quants of YuYu1015/YuYu1015-Ornith-1.0-35B-abliterated (the BF16 source) · also available as NVFP4
English
imatrix-calibrated, Unsloth-Dynamic-style GGUF quants of the abliterated (uncensored) Qwen3.5 35B MoE reasoning model. Sensitive tensors are kept at higher precision (state-space / GatedDeltaNet, router, attention, embeddings) while the routed experts carry the compression — so quality holds up far better than a flat quant.
Files
| Quant | Size | Notes |
|---|---|---|
| Q8_0 | 36.9 GB | Near-lossless, highest quality |
| UD-Q6_K | 28.8 GB | High quality — recommended if you have the VRAM/RAM |
| UD-Q4_K_M | 21.5 GB | Best size/quality balance — recommended for most |
- UD = imatrix (chat+code+wiki calibration) + per-tensor dynamic precision:
ssm_*(GatedDeltaNet) → Q8_0, router → F32, attention → Q5_K (in Q4_K_M), output/embeddings kept high; routed experts → Q4_K / Q6_K. - BF16 source: YuYu1015-Ornith-1.0-35B-abliterated · NVFP4: -NVFP4
Requirements
Use the latest llama.cpp — Qwen3.5's hybrid GatedDeltaNet / SSM layers need recent operators.
Recommended Sampling Parameters
This is a reasoning model (emits <think>…</think>). Use the official Qwen3.5 settings and keep repeat-penalty at 1.0:
--temp 1.0 --top-p 0.95 --top-k 20 --min-p 0.0 --repeat-penalty 1.0
Do not raise
repeat-penalty(truncates). For long-form, keeppresence-penalty 0.
Usage (llama.cpp)
llama-cli -m Ornith-35B-UD-Q4_K_M.gguf --temp 1.0 --top-p 0.95 --top-k 20 --repeat-penalty 1.0 -cnv
Safety Warning
This model has safety filtering removed (abliterated) and may generate sensitive or inappropriate content. Users are solely responsible for all consequences and legal liability, and must ensure usage complies with local laws and ethical standards.
Credits
- BF16 source: YuYu1015/YuYu1015-Ornith-1.0-35B-abliterated
- Base model: deepreinforce-ai/Ornith-1.0-35B
- Author: YuYu1015
繁體中文
YuYu1015/YuYu1015-Ornith-1.0-35B-abliterated(BF16 來源)的 GGUF 量化版本;另有 NVFP4 版本。
以 imatrix 高校準 + Unsloth-Dynamic 風格量化的 abliterated(去審查)Qwen3.5 35B MoE 推理模型。敏感張量保高精度(state-space / GatedDeltaNet、router、attention、embedding),壓縮集中在 routed experts —— 品質遠優於整體單一量化。
檔案
| 量化 | 大小 | 說明 |
|---|---|---|
| Q8_0 | 36.9 GB | 近乎無損,最高品質 |
| UD-Q6_K | 28.8 GB | 高品質 —— 記憶體夠建議選這個 |
| UD-Q4_K_M | 21.5 GB | 大小/品質最佳平衡 —— 多數人推薦 |
- UD = imatrix(chat+code+wiki 校準)+ 逐張量動態精度:
ssm_*(GatedDeltaNet)→ Q8_0、router → F32、attention → Q5_K(Q4_K_M 中）、output/embedding 保高;routed experts → Q4_K / Q6_K。 - BF16 來源: -abliterated · NVFP4: -NVFP4
需求
請用最新 llama.cpp —— Qwen3.5 的 GatedDeltaNet / SSM 混合層需要新算子。
建議取樣參數
這是推理模型(輸出 <think>…</think>)。用 Qwen3.5 官方設定,repeat-penalty 保持 1.0:
--temp 1.0 --top-p 0.95 --top-k 20 --min-p 0.0 --repeat-penalty 1.0
請勿調高
repeat-penalty(會截斷);長文請保持presence-penalty 0。
使用方式(llama.cpp)
llama-cli -m Ornith-35B-UD-Q4_K_M.gguf --temp 1.0 --top-p 0.95 --top-k 20 --repeat-penalty 1.0 -cnv
安全警告
此模型已移除安全過濾(abliterated),可能產生敏感或不當內容。使用者須自行承擔所有風險與法律責任,並確保使用方式符合當地法規與倫理標準。
致謝
- BF16 來源: YuYu1015/YuYu1015-Ornith-1.0-35B-abliterated
- 基礎模型: deepreinforce-ai/Ornith-1.0-35B
- 作者: YuYu1015